SRL for low resource languages isn’t needed for semantic SMT

SRL for low resource languages isn’t needed for semantic SMT Meriem Beloucif and Dekai Wu Human Language Technology Center Hong Kong University of Science and Technology Clear Water Bay, Hong Kong mbeloucif|dekai@cs.ust.hk Abstract Previous attempts at injecting semantic frame biases into SMT training for low re- source languages failed because either (a) no semantic parser is available for the low resource input language; or (b) the output English language semantic parses excise relevant parts of the alignment space too aggressively. We present the first seman- tic SMT model to succeed in significantly improving translation quality across many low resource input languages for which no automatic SRL is available —consis- tently and across all common MT met- rics. The results we report are the best by far to date for this type of approach; our analyses suggest that in general, eas- ier approaches toward including semantics in training SMT models may be more fea- sible than generally assumed even for low resource languages where semantic parsers remain scarce. While recent proposals to use the crosslin- gual evaluation metric XMEANT during inversion transduction grammar (ITG) in- duction are inapplicable to low resource languages that lack semantic parsers, we break the bottleneck via a vastly im- proved method of biasing ITG induction toward learning more semantically cor- rect alignments using the monolingual se- mantic evaluation metric MEANT. Un- like XMEANT, MEANT requires only a readily-available English (output lan- guage) semantic parser. The advances we report here exploit the novel realiza- tion that MEANT represents an excel- © 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. lent way to semantically bias expectation- maximization induction even for low re- source languages. We test our systems on challenging languages including Amharic, Uyghur, Tigrinya and Oromo. Results show that our model influences the learn- ing towards more semantically correct alignments, leading to better translation quality than both the standard ITG or GIZA++ based SMT training models on different datasets. 1 Introduction Statistical machine translation (SMT) for low resource languages has been a difficult task due to the unavailability of large parallel corpora. It becomes imperative to make learning from small data more efficient by adding additional constraints to create stronger inductive biases—especially lin- guistically well-motivated constraints, such as the shallow semantic parses of the training sentences. However, while automatic semantic role labeling (SRL) is readily available to produce shallow se- mantic parses for a high-resource output language (typically English), the problem is that SRL is usu- ally not available for low resource input languages such as Tigrinya, Oromo, Uyghur or Uzbek. In this paper, we propose a new method which adopts the monolingual semantic evaluation met- ric MEANT as a confidence-weighting measure to assess the degree of goodness of training in- stances, giving a newer strategy than Beloucif and Wu (2016a) who used the degree of compatibil- ity or similarity between the semantic role label- ing of the input and output sentences. Their ap- proach might …

Content truncated to 500 words